Adversarial code review for GitHub PRs with:
- Cynical internal review phase, professional external output
- Severity-based findings with [likely] tag for high-confidence issues
- LLM disclaimer in output footer
- Safety features: sandboxed execution, large PR warnings, preview before post